Replies: 1 comment
-
|
Whaha, I think I just rubber-ducked myself by making this post. I needed to specify a stop-sequence, I added ### in between each example and set the stop sequence to ###. It now stops when it should. I do still have the problem that it misunderstands intention, for instance (I shortened the prompt to improve performance) And it outputs this: It somehow decided that the "buy chicken" is finished the moment it's added.. hmm |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm experimenting a bit with LLaMa to make a bot that can help with making todolists and help you focus. Mostly aimed at people with executive function problems (like myself, hah). Nothing serious just yet, but I do see some potential here for a sort of personalized executive-functioning coach if I can get this to work well..
I'm using langchain to do this. One task that I need accomplished is that once the user made a decision about how to change the todolist, the bot should update it.
Here's the prompt I'm using:
When using GPT-4, I can use an instruction-prompt which has very little context and performs very consistently.
With LLaMa, I'm struggling to get the right output with only the instructions -- which makes sense, since it's not finetuned --, so I thought few-shot would be the best alternative. I also read this prompt engineering post to get some ideas
The kind of output I'm getting doesn't make a lot of sense though, it's just making up more examples instead of actually performing the task
Here's one test task I'm doing:
Output is:
It doesn't really matter if I pick a large or small model, they all have the same problem. The annoying part is that what it outputs in the beginning is actually correct and desired, it modified the initial todolist correctly with the new tasks. It's just that the rest of the output is undesired.
Can anyone guide me in the right direction? Thanks!
Beta Was this translation helpful? Give feedback.
All reactions